reimplemented model downloading

on_server_start.py

@@ -1,11 +1,7 @@
 """
 A script that is run when the server starts.
 """
-import os
-
-from transformers import AutoModelForCausalLM
-
-DOWNLOADED_MODELS_DIR = os.path.join(os.path.dirname(__file__), 'models')
+from huggingface_hub import snapshot_download
 
 
 def download_model(model_name: str):
@@ -13,9 +9,12 @@ def download_model(model_name: str):
     Downloads a model from hugging face hub to the disk but not to the RAM.
     :param model_name: The name of the model to download.
     """
-    print(f"Downloading model: {model_name}")
-    AutoModelForCausalLM.from_pretrained(model_name)
-    print(f"Downloaded model: {model_name}")
+    number_of_seconds_in_a_day: int = 86_400
+    snapshot_download(
+        repo_id=model_name,
+        etag_timeout=number_of_seconds_in_a_day,
+        resume_download=True,
+    )
 
 
 def download_useful_models():
@@ -24,7 +23,6 @@ def download_useful_models():
     So that the user doesn't have to wait for the models to download when they first use the app.
     """
     print("Downloading useful models...")
-    os.makedirs(DOWNLOADED_MODELS_DIR, exist_ok=True)
     useful_models = (
         "facebook/opt-125m",
     )